Research work · 2025
Development and Assessment of Software-Level Optimization Techniques to Enhance the Energy Efficiency of Spiking Neural Networks
Jayakumar Rajarajeswaran
Abstract
Spiking Neural Networks (SNNs) offer significant advantages for energy-efficient computing due to their event-driven processing. However, the computational demands of training and inference in SNNs remain a challenge, leading to high energy consumption that limits their deployment in real-world applications. This study introduces software-level optimizations aimed at improving energy efficiency without compromising accuracy. The proposed techniques include algorithmic refinements that minimize redundant operations, dynamic precision scaling that adjusts computation accuracy based on neural activity and adaptive event scheduling that optimizes spike propagation. Experimental results demonstrate that these techniques collectively achieve a 35% reduction in inference energy and a 28% reduction in training energy, while maintaining accuracy loss within 2%. The findings provide a pathway for deploying SNNs in power-constrained environments such as edge devices and neuromorphic processors, ensuring both computational efficiency and sustainability.